You can believe (like I do) that LLMs are a genuine breakthrough for bringing code and knowledge retrieval to a wider audience while recognizing that the AI infrastructure buildout has detached from economic reality.
Here’s the basic argument:
Interest rates are rising, in part due to competition for capital and in part due to the Iran War making everything more expensive.
AI data centers and the GPUs to fill them are primarily being funded by increasingly expensive debt.
US-based AI labs like Anthropic, OpenAI, Google Deepmind, and Meta Superintelligence Lab are the primary buyers of the new AI infrastructure that is being built, and their financial projections are the justification behind the scale of that build-out.
Rather than living up to the rosy projections, revenue for US-based AI labs is plateauing and their margins are getting competed away by cheap open weights competitors like DeepSeek, Z.ai, and Moonshot.
Put those four things together, and the inevitable conclusion is that the companies borrowing half a trillion dollars at rising interest rates to build GPU data centers are relying on customers who won’t be able to pay their tab.
Let’s dig in.
For the first couple of years of the AI boom, Big Tech paid for its GPU purchases using cash from existing lines of business like search, ads, and cloud compute. That phase is over.
In 2026, AI-related debt issuance has exploded more than 4x to nearly $500 billion. Alphabet, Amazon, Meta, and Oracle alone issued more than $228 billion in bonds by mid-October 2026, more than double their total for all of 2025. JPMorgan now expects the big cloud companies to spend $800 billion this year and over $1 trillion a year by 2027, most of it financed with debt.
You don’t need to be levered to the hilt for this math to go sideways. Even if you are a massive company with multiple revenue streams and won’t go bankrupt any time soon, borrowing long-term money at rising interest rates to buy GPUs that depreciate in half a decade or so only makes sense if the revenue those chips generate grows fast enough to pay for that debt. Otherwise, every newly borrowed dollar destroys value for your shareholders, and eventually either your CFO or your shareholders will force you to stop.
For example, Oracle:
It’s already carrying $125.3 billion in debt, more than three times its $37.1 billion in cash.
The interest it’s paying every quarter just increased 55% year-over-year to $1.43 billion.
On top of that, it has signed $288 billion in new long-term data center leases that haven’t even shown up on its balance sheet yet.
If you want to see what the unit economics of a GPU data center look like without an attached software monopoly to finance the losses, look at the canary in the coal mine: CoreWeave.
CoreWeave doesn’t have any problems with demand: its capacity is effectively sold out, it has a backlog of over $100 billion of signed contracts, and its revenue more than doubled last quarter to $2.6 billion. Yet the bond market is suddenly treating it like a distressed asset:
To build out that capacity, CoreWeave racked up $35.6 billion in debt (an 8.94x debt-to-equity ratio, rated B+ junk).
In Q2 2026, it made $128 million in adjusted operating income while paying $640 million in interest. For Q3, it expects $200M–260M in operating profit against $860M–950M in interest expense, meaning interest alone eats more than a quarter of its revenue and roughly 4x what the business actually earns.
In mid-October, yields on CoreWeave’s bonds blew out to roughly 16%, and its five-year credit default swaps topped 11.4%, well past the 10% threshold that is generally interpreted to signal severe credit distress.
CoreWeave isn’t an isolated case; it’s just the pure-play version of the exact same cost-of-capital math every data center operator is facing now. Of course, unlike CoreWeave, Big Tech won’t default on their loans any time soon. They’ll just be forced to stop increasing their GPU orders by the return-on-capital math.
Why haven’t chipmakers like Nvidia or the broader semiconductor supply chain felt any of this pain yet? Because the data center operators raise debt and pay Nvidia up front when the chips are delivered. Nvidia’s quarterly numbers look great right up until the bond market cuts off its customers’ ability to pay.
The entire justification for borrowing all that money to build GPU data centers is the assumption that AI lab revenue will keep compounding at 100%+ per quarter for years to come.
Instead, if you look closely at the numbers coming out of Anthropic’s pre-IPO disclosures this month, the growth curve is looking pretty flat.
It’s true that Anthropic’s quarterly revenue has grown massively this year, from $4.7 billion in Q1 2026 to $11.6 billion in Q2 (+147% quarter-over-quarter) to $16.8 billion in Q3 (+45% quarter-over-quarter). Look at the monthly math inside Q3, however, and you can see why investors at its October pre-IPO meeting got nervous:
By the end of July 2026, Anthropic had already hit a $65 billion annualized revenue run rate ($5.42 billion a month).
As PitchBook pointed out, if Anthropic’s revenue had simply stopped growing altogether after July, a completely flat Q3 would have come in at $16.25 billion.
Instead, actual Q3 revenue came in at $16.8 billion ($67.2 billion annualized)—meaning across August and September, Anthropic’s run rate grew barely 3.4% over its July pace.
Why is revenue stalling out so abruptly? A few reasons.
One component is that a huge share of AI lab revenue isn’t self-sustaining: it is exogenous.
Healthy software revenue is endogenous: a business buys software, that software helps them make more money or cut real costs, and they use a slice of those ongoing profits to keep paying for more of the software. A significant portion of AI token revenue, by contrast, is being pumped in from the outside by finite pools of venture capital.
Corporate credit-card and billing data from Ramp (highlighted this week by Apollo Chief Economist Torsten Sløk) shows that among the top 1% of AI token buyers, 80.4% are VC-backed startups (compared to just 13.1% for the median token buyer). Essentially, VC investments into AI wrapper startups are a big driver of API token revenue.
And even for the developers and businesses who are getting real, everyday value out of LLMs, US labs are watching their pricing power get eroded by much cheaper open-weight models from Chinese companies like DeepSeek, Z.ai, and Moonshot.
Take my own workflow: I use LLMs to generate glue code regularly, and I recently canceled my Claude subscription to run the new DeepSeek-V4.1-Flash-1017 inside a confidential compute environment. For most daily tasks, it’s roughly as good as Claude, activates only 16B parameters per output token, and costs roughly 1/20th as much compute per task. All that while giving me better privacy because my code stays inside an encrypted hardware enclave instead of sitting in plaintext on Anthropic’s servers.
Whenever you bring up this kind of price deflation, AI bulls like to shout “Jevons Paradox!” and claim that if tokens get 20x cheaper, people will use at least 20x more of them.
I do think there is a real but moderate Jevons effect here: because inference is so cheap now, I’m much more likely to implement marginal automations than I would have been before. But it isn’t unbounded. In practice, it’s maybe 2x or 3x as much code as I would have wanted at the old price, because I still have to specify what to build, review the output, and maintain it. When the price per task drops 20x, and my usage only goes up 2x to 3x, total dollar revenue into the ecosystem (and net GPU demand per user) still collapses by 70% to 85%.
What about test-time compute? Isn’t the opportunity to scale that up going to keep demand steady as token prices drop? No, because the quality of the output is logarithmic with the amount of test-time compute, so more test-time compute has diminishing returns.
If the financial math is this questionable, why hasn’t the tech industry noticed?
Over the last year and a half, LLMs have gotten good enough at writing code that most software developers use them as part of their workflows. That happened for a specific technical reason: code is an abstract domain where you can run automated tests to verify correctness, which allowed labs to use Reinforcement Learning from Verifiable Rewards (RLVR) to make models dramatically better at generating useful code. And it continues to improve as more training data is synthesized.
For the tech industry, writing code is the main bottleneck of the business. So when software engineers and tech executives saw how much faster they could produce code, they projected their own daily experience onto the entire global economy.
Everywhere else, however, code is only a piece of the puzzle. Most other parts of the economy, like hospitals, logistics, manufacturing, construction, retail, and energy, have a lot of other inputs besides code, and real-world business problems don’t come with a clean reward signal that lets you optimize an AI model to solve them.
You can see this split clearly in the macro data published this week by Apollo Chief Economist Torsten Sløk:
At the “Magnificent 7” tech giants (where software is the core product), inflation-adjusted revenue per employee has surged to roughly $780,000, and profit margins have climbed toward 30%.
Across the S&P 493 (the rest of the S&P 500), inflation-adjusted revenue per employee has flatlined at $525,000, and profit margins haven’t budged from 10%.
As Sløk put it, there is still zero financial payoff from AI outside the tech sector. And you cannot pay back trillions of dollars of high-yield data center debt just by selling developer tools.
Bulls love to argue that legacy enterprises just take three to five years to adopt new technology, like they did with cloud migration in the 2010s. They forget that if it takes five years for the S&P 493 to figure out how to use these tools, the GPUs being bought with debt today will already be essentially obsolete.
It’s worth stepping back and asking how some of the smartest guys in the room talked themselves into borrowing so much money to fund data centers without checking whether end users would actually pay for them.
The short answer is a toxic combination of ideological delusion at the AI labs and a classic corporate Prisoner’s Dilemma at the tech giants. At the labs, the buildout is driven by LessWrong rationalists who convinced themselves that we are on the cusp of Artificial Superintelligence (ASI), a digital god that will usher in a new world order and render normal unit economics irrelevant. If you truly believe you are two years away from building a machine that replaces all human labor on earth, signing $750 billion in compute commitments on $16.8 billion of quarterly revenue feels totally rational. Meanwhile, hard-nosed executives at Oracle, Amazon, Google, and Meta don’t even have to believe in the ASI cult; they are trapped in a FOMO arms race where burning tens of billions on terrible ROI feels safer than risking market share to an upstart or a rival.
That’s where the gap between Silicon Valley’s compute commitments and the rest of the economy comes from. And every month, the evidence keeps stacking up against them. We didn’t summon a superintelligent god or a replacement for a company’s workforce; we built a very useful, steadily improving tool for bringing code, math, and knowledge retrieval to a wider audience.
And so we are staring at a massive infrastructure overbuild, and when it unwinds, it is going to cause a sizable stock market downturn that will take a big chunk out of your 401(k). Looking at how the acceleration of end-customer demand has already peaked, that may be as soon as late 2026 or early 2027. Perhaps triggered by the rumored upcoming Anthropic IPO.
I’m not recommending anyone follow my lead, but once the music stops on the data center financing, the AI chip designers and the rest of the semiconductor supply chain are next in line (and, standing on the wrong side of the bullwhip effect, they’ll feel the financial impact even harder). I loaded up on long-dated SOXX puts after CoreWeave started showing signs of distress a couple weeks ago, because I don’t want to go down with this ship.
Aside: I chose SOXX because I don’t have the time or the deep AI infrastructure insider knowledge to try and tease apart whether the capex freeze hits GPUs, custom silicon providers, high-bandwidth memory, or fab equipment hardest. It’s a good balance across all of them.
And another little note: I realize that Google has their own in-house AI accelerator silicon and they call them TPUs. Cerebras and others have their own names for their AI accelerators. But I didn’t want my writing to get bogged down with all those details, and it doesn’t change anything about the reasoning above.